Re: [PATCH v5 8/9] vfs: Add vfs_copy_file_range() support for pagecache copies

On Wed, Oct 14, 2015 at 01:59:40PM -0400, Anna Schumaker wrote:

On 10/12/2015 07:17 PM, Darrick J. Wong wrote:

quoted

On Sun, Oct 11, 2015 at 07:22:03AM -0700, Christoph Hellwig wrote:

quoted

On Wed, Sep 30, 2015 at 01:26:52PM -0400, Anna Schumaker wrote:

quoted

This allows us to have an in-kernel copy mechanism that avoids frequent
switches between kernel and user space.  This is especially useful so
NFSD can support server-side copies.

I make pagecache copies configurable by adding three new (exclusive)
flags:
- COPY_FR_REFLINK tells vfs_copy_file_range() to only create a reflink.
- COPY_FR_COPY does a full data copy, but may be filesystem accelerated.
- COPY_FR_DEDUP creates a reflink, but only if the contents of both
  ranges are identical.

All but FR_COPY really should be a separate system call.  Clones (an
dedup as a special case of clones) are really a separate beast from file
copies.

If I want to clone a file I either want it clone fully or fail, not copy
a certain amount.  That means that a) we need to return an error not
short "write", and b) locking impementations are important - we need to
prevent other applications from racing with our clone even if it is
large, while to get these semantics for the possible short returning
file copy will require a proper userland locking protocol. Last but not
least file copies need to be interruptible while clones should be not.
All this is already important for local file systems and even more
important for NFS exporting.

So I'd suggest to drop this patch and just let your syscall handle
actualy copies with all their horrors.  We can go with Peng's patches
to generalize the btrfs ioctls for clones for now which is what everyone
already uses anyway, and then add a separate sys_file_clone later.

So what I'm hearing is that I should drop the reflink and dedup flags and
change this system call only perform a full copy (with preserving of
sparseness), correct?  I can make those changes, but only if everybody is in
agreement that it's the best way forward.

Sounds fine to me; I'll work on promoting EXTENT_SAME to the VFS.

The only reason I haven't done anything to make this system call
interruptible is because I haven't been able to find any documentation or
examples for making system calls interruptible.  How do I do this?

I thought it was mostly a matter of sprinkling in "if (signal_pending(...))
return -ERESTARTSYS" type things whenever it's convenient to check.  The splice
code already seems to have this, though I'm no expert on what the splice code
actually does. :)

--D

Anna

quoted

Hm.  Peng's patches only generalize the CLONE and CLONE_RANGE ioctls from
btrfs, however they don't port over the (vastly different) EXTENT_SAME ioctl.

What does everyone think about generalizing EXTENT_SAME?  The interface enables
one to ask the kernel to dedupe multiple file ranges in a single call.  That's
more complex than what I was proposing with COPY_FR_DEDUP(E), but I'm assuming
that the extra complexity buys us the ability to ... multi-dedupe at the same
time, with locks held on the source file?

I'm happy to generalize the existing EXTENT_SAME, but please yell if you really
hate the interface.

--D

quoted

--
To unsubscribe from this list: send the line "unsubscribe linux-api" in
the body of a message to majordomo-u79uwXL29TY76Z2rM5mHXA@public.gmane.org
More majordomo info at  http://vger.kernel.org/majordomo-info.html

--
To unsubscribe from this list: send the line "unsubscribe linux-nfs" in
the body of a message to majordomo-u79uwXL29TY76Z2rM5mHXA@public.gmane.org
More majordomo info at  http://vger.kernel.org/majordomo-info.html

On Wed, Oct 14, 2015 at 11:11 AM, Darrick J. Wong
[off-list ref] wrote:

On Wed, Oct 14, 2015 at 01:59:40PM -0400, Anna Schumaker wrote:

quoted

On 10/12/2015 07:17 PM, Darrick J. Wong wrote:

quoted

On Sun, Oct 11, 2015 at 07:22:03AM -0700, Christoph Hellwig wrote:

quoted

On Wed, Sep 30, 2015 at 01:26:52PM -0400, Anna Schumaker wrote:

quoted

This allows us to have an in-kernel copy mechanism that avoids frequent
switches between kernel and user space.  This is especially useful so
NFSD can support server-side copies.

I make pagecache copies configurable by adding three new (exclusive)
flags:
- COPY_FR_REFLINK tells vfs_copy_file_range() to only create a reflink.
- COPY_FR_COPY does a full data copy, but may be filesystem accelerated.
- COPY_FR_DEDUP creates a reflink, but only if the contents of both
  ranges are identical.

All but FR_COPY really should be a separate system call.  Clones (an
dedup as a special case of clones) are really a separate beast from file
copies.

If I want to clone a file I either want it clone fully or fail, not copy
a certain amount.  That means that a) we need to return an error not
short "write", and b) locking impementations are important - we need to
prevent other applications from racing with our clone even if it is
large, while to get these semantics for the possible short returning
file copy will require a proper userland locking protocol. Last but not
least file copies need to be interruptible while clones should be not.
All this is already important for local file systems and even more
important for NFS exporting.

So I'd suggest to drop this patch and just let your syscall handle
actualy copies with all their horrors.  We can go with Peng's patches
to generalize the btrfs ioctls for clones for now which is what everyone
already uses anyway, and then add a separate sys_file_clone later.

So what I'm hearing is that I should drop the reflink and dedup flags and
change this system call only perform a full copy (with preserving of
sparseness), correct?  I can make those changes, but only if everybody is in
agreement that it's the best way forward.

Sounds fine to me; I'll work on promoting EXTENT_SAME to the VFS.

quoted

The only reason I haven't done anything to make this system call
interruptible is because I haven't been able to find any documentation or
examples for making system calls interruptible.  How do I do this?

I thought it was mostly a matter of sprinkling in "if (signal_pending(...))
return -ERESTARTSYS" type things whenever it's convenient to check.  The splice
code already seems to have this, though I'm no expert on what the splice code
actually does. :)

Oh, right.  That's for making loops that don't otherwise block
interruptible.  If you're doing wait_xyz, then you want to use the
interruptable variable of that.

Anyway, I just checked on x86.  The relevant error codes are (I think):

-EINTR: returns -EINTR to userspace with no special handling.

-ERESTARTNOINTR: end the syscall, call a signal handler if
appropriate, then retry the syscall with the same arguments (i.e. the
syscall needs to make sure that trying again is an acceptable thing to
do by, e.g., updating offsets that are used).

-ERESTARTSYS: same as -ERESTARTNOINTR *unless* there's an unblocked
signal handler that has SA_RESTART clear, which which case the caller
gets -EINTR.

-ERESTARTNOHAND: end the syscall and retry with the same arguments if
no signal handler would be called; otherwise call the signal handler
and return -EINTR to the caller.

-ERESTART_RESTARTBLOCK: return -EINTR if a signal is delivered and
otherwise use the restart_block mechanism.  (Don't use that -- it's
evil.)

So -ERESTARTSYS is probably the most sensible thing to use under
normal circumstances.

--Andy